Daily papers

Wednesday, 11 November 2026 · 6 min read

APEX improves streaming video generation efficiency and quality by sharing fake estimation within the generator

APEX extends condition shifting to streaming video, removing the need for a separate fake-score network and reducing training cost. It achieves higher instruction fulfillment and visual aesthetics than S2-Avatar, with positive human preference and lower GPU memory and iteration time than DMD.

Based on the paper Few-Step Streaming Video Generation via Condition Shifting by Deyuan Liu, Jungang Li, Kaiwen Zheng, Kai Jiang and 7 others

Challenges in efficient streaming video generation and distribution matching

Streaming video generation requires producing long sequences efficiently, which is costly when using diffusion or flow models that need multiple network evaluations per video block.

Few-step sampling reduces computation by limiting evaluations, but maintaining visual quality, character consistency, and responsiveness to instructions remains difficult.

Distribution Matching Distillation (DMD) supervises few-step generators using a separate fake-score network to estimate the generated distribution, but this adds memory and computation overhead and requires careful coordination with the evolving generator.

Sign in to read the whole paper

You have read the opening. The rest is free with a Bezen account, and no card is needed.

In the rest of this paper

  • How the method works, step by step, with a flow chart
  • Three key formulas from the paper, explained symbol by symbol
  • Three figures from the paper, and how to read them
  • How it was tested, what was found, and where it falls short
APEX improves streaming video generation efficiency and quality by sharing fake estimation within the generator · Bezen